Download Predicting the Genes Regulated by MicroRNAs via Binding Sites in

Computational Chemistry in switzerland
doi:10.2533/chimia.2014.629
CHIMIA 2014, 68, Nr. 9
Chimia 68 (2014) 629–632
629
© Schweizerische Chemische Gesellschaft
Predicting the Genes Regulated by
MicroRNAs via Binding Sites in the
3' Untranslated and Coding Regions
Jiří Vaníček*
Abstract: MicroRNAs form one of the groups of small noncoding RNA molecules that have completely changed
our understanding of gene regulatory networks. Because microRNAs have been discovered only relatively
recently, most of their functions remain unknown, providing a challenge to both experiment and theory. I review
several computational approaches pursued in our group to answer this challenge. In particular, I show that a few
rather simple ideas can go a long way in predicting accurately genes regulated by microRNAs via binding sites
both in the coding and 3' untranslated regions (3'UTRs). Finally, I mention briefly several applications, including
two collaborations with experimental groups, which have shed new light on the latency and reactivation of
herpesviruses, and on the maturation of red blood cells.
Keywords: Conservation · MicroRNA · MicroRNA target · Target prediction algorithm ·
RNA secondary structure
1. Introduction
MicroRNAs (miRNAs) are small, 20–
23 nucleotides (nt) long noncoding RNA
molecules that function as negative gene
regulators by binding to target messenger RNAs (mRNAs) and either degrading
them or repressing their translation into
protein.[1] MicroRNAs play an important
role in cell differentiation, development,
cancer, and other biological processes in
species ranging from viruses to humans.[2]
While more than thousand miRNAs are encoded in the human genome, most of their
target genes remain unknown, especially
since direct validation of the functionality
of individual miRNA-target pairs is rather
tedious and expensive.
Fortunately, several high-throughput
techniques can be used for indirect validation, including high-throughput proteomics
methods,[3,4] in which protein levels of
thousands of genes are monitored upon
overexpression of a miRNA of interest,
*Correspondence: Prof. Dr. J. Vaníček
Laboratory of Theoretical Physical Chemistry
Institut des Sciences et Ingénierie Chimiques
Ecole Polytechnique Fédérale de Lausanne (EPFL)
Avenue Forel, CH-1015 Lausanne
Tel.: +41 21 693 47 36
E-mail: [email protected]
various methods for quantifying mRNA
expression (microarrays, RNA-Seq, etc.),
or high-throughput cross-linking immunoprecipitation (CLIP),[5,6] which provides coordinates of thousands of mRNA
regions bound by the Argonaute-miRNA
ribonucleoprotein complexes. However,
since these techniques are indirect and
even more expensive than direct validation, they are usually combined with computational approaches to focus the search
for functional miRNA targets, and this is,
in fact, one of the research interests in our
group.
2. Predicting MicroRNA Targets in
the 3'Untranslated Region (3'UTR)
The challenge faced by miRNA target
prediction algorithms becomes obvious
by comparing miRNAs with closely related small interfering RNAs (siRNAs):
Predicting targets of siRNAs is relatively
straightforward since these ~22 nt long
RNA molecules must be fully complementary to their target RNA sequences;
one can simply scan the human genome
for such complementary sequences and be
rather confident that those found will be
functional. Indeed, assuming that each of
the four (A, C, G, T) nucleotides is equally
likely and ignoring genomic repeats, one
finds that a specific sequence of 22 nt
would appear by chance on average only
once in a sequence of 422 ≈ 1.8 × 1013 nt.
In other words, it would rarely show up by
chance in the human genome whose length
is approximately 3 × 109 nt.
Searching for targets of miRNAs, which
are also ~22 nt long, is more difficult since
they require complementarity only within
a so-called ‘seed’ region, i.e. a region of
six to seven consecutive nucleotides starting at position 2 of the miRNA (see Fig. 1).
Assuming a typical human 3'UTR length
of about 1000 nt, a simple back-of-the-envelope calculation shows that the naïve algorithm requiring complementarity to the
6-mer seed would predict that the 3'UTR
of every fourth gene is regulated by each
miRNA. The algorithm that works so well
for siRNAs fails for miRNAs; it is therefore necessary to use additional criteria to
predict miRNA targets precisely.
3
5
mRNA
seed
5
m
iR
N
A
3
Fig. 1. Simplified representation of the binding
between a miRNA and its mRNA target. The
most important interaction occurs along a seed
region of six to seven nucleotides.
Among the many criteria used to improve upon the naïve algorithm are conservation and accessibility of the binding site,
hybridization energy between the miRNA
and target mRNA, and various empirical
rules based on training sets constructed
630
from experimentally validated miRNAtarget pairs. It is not surprising that conservation of the binding site among several
species (Fig. 2a) has been used successfully in many miRNA target prediction algorithms.[7–13] The justification is simple –
sequences that have been conserved during
evolution are much more likely to be functional. Another criterion uses the accessibility of the binding site and is based on the
assumption that miRNAs are more likely
to bind to accessible segments of mRNA,
i.e. single-stranded regions of the secondary RNA structure (Fig. 2b). However, it
is not necessary that the full ‘seed match’
(i.e. mRNA sequence complementary to
the seed) be accessible at all times; four
nucleotides are often sufficient to nucleate the binding (see Section 4 for more
details).[14] Interestingly, in one of our papers,[15] we showed that some ‘obviously
good’ criteria such as the hybridization
energy between the miRNA and mRNA
are poor for ranking predictions. One may
object that the free energy of the interaction should also include the ‘opening energy’, i.e. the energy required to unwind
the secondary mRNA structure to make
it accessible for binding by the miRNA.
Yet, we showed that even the ‘total’ free
energy, computed as a signed sum of the
opening and hybridization energies does
not significantly improve the precision
of target predictions beyond the simple
algorithm requiring complementarity to a
7-mer seed.[15]
As a consequence, thermodynamic criteria are not used for ranking predictions in
our algorithms.[13,15,16] Instead, we rank the
predictions according to the over-representation of the seed match compared to a random background.[12] This is motivated by
the assumption that functionally interacting miRNA-mRNA pairs have co-evolved,
implying that the number of complementary sites present in regulated genes should
be higher than the corresponding number
appearing by chance in unregulated genes.
a)
Computational Chemistry in switzerland
CHIMIA 2014, 68, Nr. 9
In the 3'UTR, which does not contain as
much information as the coding region,
we model this random background by the
second-order Markov model based on the
nucleotide and dinucleotide composition
of the given 3'UTR sequence.[12] The dinucleotide composition is important since
adjacent CpG (-cytosine-phosphate-guanine-) pairs are under-represented in comparison with what would be expected from
the C and G content alone.
In addition to this ranking criterion, we
also use conservation and accessibility of
the binding site. However, in contrast to
algorithms that rank their predictions according to the extent of conservation and
accessibility, these two criteria are in our
algorithms only used as filters – we require
a minimum amount of conservation and
accessibility, but then rank the predictions
according to over-representation.
To summarize, for each miRNA-3'UTR
pair, let us denote by c the number of conserved and accessible n-mer seed matches
in the 3'UTR and by l the total number of
conserved and accessible n-mers in the
3'UTR. We first use the Markov model to
compute the probability p to find an n-mer
seed match by chance at any particular position of a random 3'UTR,[12,16] and then
evaluate the probability PSH (the singlehypothesis P-value) to find, by chance,
at least c conserved and accessible seed
matches in a random 3'UTR containing l
conserved and accessible n-mers:
PSH =
l − n +1
⎛ l − n + 1⎞ i
l − n +1−i
(1)
⎟ p (1 − p )
⎝ i ⎠
∑ ⎜
i =c
Lower PSH values (i.e. stronger overrepresentation) imply a higher likelihood
of co-evolution between the miRNA and
the seed match, and hence, a higher likelihood of biological functionality. (In the
definition of PSH, ‘accessible site’ stands,
more precisely, for what we call a ‘partially accessible site’, i.e. an n-mer contain-
ing at least one 4-mer that appears in the
single-stranded region of at least 20% of
secondary structures from the Boltzmann
ensemble of secondary structures of the
given 3'UTR.[15] Similarly, ‘conserved
n-mer’ typically stands for an n-mer that
is conserved in the human, chimp, rhesus,
and mouse.[13])
There are several accurate prediction algorithms for predicting conserved
targets of conserved miRNAs; these include, e.g. DIANA-microT,[7] PicTar,[8]
TargetScan,[9,10] ElMMo,[11] the algorithm
of Robins and Press,[12] and our algorithm
PACCMIT.[13] However, we have shown[15]
that as soon as the conservation requirement is dropped, e.g. when one is interested in species-specific targets, the precision
of existing algorithms decreases drastically. Yet, we have shown that our simple
algorithm based on accessibility and overrepresentation (PACMIT)[15] for predicting
nonconserved targets remains accurate in
such situations.[13] The algorithm was validated[13,15] on four very different and complementary datasets: the high-throughput
(1) mRNA expression and (2) proteomics
datasets of Selbach et al.[3] and Baek et
al.,[4] in which protein and RNA levels of
several thousand genes were monitored
upon overexpression (or knockdown) of
several miRNAs, (3) the high-throughput
cross-linking immunoprecipitation (CLIP)
datasets of Hafner et al.[5] and Chi et al.,[6]
providing coordinates of thousands of
mRNA regions bound by the ArgonautemiRNA ribonucleoprotein complexes, and
(4) a dataset of Kertesz et al.[17] using luciferase reporter assays.
3. Predicting MicroRNA Targets in
the Coding Region
Predicting miRNA targets in the coding
region is much more difficult than in the
3'UTR for an obvious reason – the coding region contains another, biologically
b)
Fig. 2. (a) Conservation of miRNA binding site among several species. (b) Accessibility of miRNA binding site in the target mRNA.
Computational Chemistry in switzerland
much more important signal – the code for
the protein. In other words, the signal that
we seek (miRNA binding sites) is much
weaker than the background (code for the
protein); it is like looking for a needle in a
haystack.
To predict miRNA targets in the coding region, we employ the same main idea
as before – we look for binding sites that
are over-represented compared to a random background, but now the background
is constructed differently. The random
background should not destroy the most
important function of the coding region –
the code for the protein. Fortunately, due
to the redundancy of the genetic code such
a background exists: since there are 43 =
64 codons (nucleotide 3-mers) and only
20 amino acids, there exist synonymous
codons encoding the same amino acid.
The desired background is constructed by
replacing each codon in the real sequence
with a randomly chosen synonymous codon. A less known fact adds another constraint to the choice of an optimal background. Namely, codon usage differs in
different species and sometimes even in
different genes in a single species, in order
to fine-tune translational efficiency. In order to construct a background preserving
only the codon usage, one would shuffle
the codons of each coding sequence; the
background satisfying the two constraints
simultaneously, i.e. preserving both the
amino acid sequence and codon usage, is
constructed by shuffling only the synonymous codons (see Fig. 3).[18,19]
To summarize, let c again denote the
number of seed matches of a given miRNA
in the coding sequence of a given mRNA.
In our PACCMIT-CDS algorithm,[20] the
miRNA-mRNA pairs are ranked according
to the probability that at least c seed matches would appear in the randomly generated
coding sequence, in practice computed as
the single-hypothesis P-value
PSH =
Nc
,
N total
(2)
where Nc is the number of random sequences with at least c seed matches and Ntotal
is the total number of random sequences.
The random sequences are generated with
the Durstenfeld modification of the FisherYates algorithm for generating random
permutations (‘shuffles’) of an array with
N elements.[21] While the P-values of the
top predictions are below 10–8, evaluating
all P-values for the whole genome with
this resolution would take hundreds of
years on a supercomputer, and, moreover,
is unnecessary. Instead, the calculation
was enormously accelerated by a gradual
refinement of P-values: the high P-values
were only evaluated with resolution 10–3,
CHIMIA 2014, 68, Nr. 9
631
Fig. 3. Construction of a random background sequence required for
predicting miRNA targets in the coding region with PACCMIT-CDS. The
background is obtained by shuffling synonymous codons in order to preserve both the amino acid sequence encoded by the gene and the codon
usage required for efficient translation.
and lower P-values for more significant
miRNA-mRNA pairs with higher resolution, as needed.
By analyzing the coding sequences of
the human genome, we demonstrated that
the background preserving both the amino
acid sequence and codon usage reduces the
noise in target predictions more than backgrounds preserving none or only one of
the two constraints. Moreover, we showed
that considering conservation of the seed
matches among related species increases
enormously the signal-to-noise of the predictions. Note that the algorithm as well as
the above analysis, including estimates of
the signal-to-noise ratio, require only the
well-known coding sequences, but – unlike other algorithms – PACCMIT-CDS
does not rely on any training sets using
proteomics, mRNA expression, or other
experimental data.
We have, nevertheless, again validated the predictions of PACCMIT-CDS
independently, in two different ways, using the large throughput (1) proteomics
dataset of Selbach et al.[3] and (2) crosslinking immunoprecipitation (CLIP) dataset of Hafner et al.[5] In order to separate
the well-established miRNA function via
binding within the 3'UTR, we had to construct datasets with binding sites with seed
matches only within the coding region.
The proteomics and cross-linking immunoprecipitation datasets confirmed that
the top predictions of PACCMIT-CDS are
a)
3
genes with more downregulated protein
levels and genes with more sites bound
by the Argonaute-miRNA ribonucleoprotein complexes within the coding region,
respectively.
4. Nucleation of MicroRNA–Target
Binding
It is agreed that the seed region is the
most important determinant of the binding between the miRNA and its target.
However, it is not clear how this seed binding is nucleated. Does the binding start
from the 3' or 5' end of the miRNA? (Fig.
4). Ray Marín, a former Ph.D. student in my
group, answered[22] this question in silico
by a careful analysis of the high-throughput cross-linking immunoprecipitation
datasets.[5,6] In particular, he compared the
accessibility of RNA segments of length 1
to 7 nucleotides at various locations within
bound seed matches with the accessibility
of the corresponding regions in unbound
seed matches. The conclusion of this analysis was that in contrast to unbound sites,
in bound sites the 3' end of the seed match
is much more accessible than the 5' end, a
result that was recently confirmed experimentally by Wan et al.[23] The miRNAmRNA binding is therefore much more
likely to start from the 5' end of the miRNA
seed, or, equivalently, from the 3' end of
the seed match within the target mRNA.[22]
b)
5
3
5
mRNA
seed
5
mRNA
seed
m
iR
N
A
5
3
m
iR
N
A
3
Fig. 4. Two alternative hypotheses about the nucleation of the miRNA-mRNA binding. (a) Binding
starts at the 5' end of the seed (corresponding to the 3' end of the seed match). (b) Binding starts
at the 3' end of the seed (corresponding to the 5' end of the seed match). By analyzing the accessibility of binding sites detected in cross-linking immunoprecipitation (CLIP) experiments, we
showed that mechanism (a) is much more likely.
632
Computational Chemistry in switzerland
CHIMIA 2014, 68, Nr. 9
5. MicroRNA Functions in Herpesvirus Infection and in the Maturation
of Red Blood Cells
In 2004 miRNAs were discovered
in herpesviruses. This was very exciting
because viruses, despite their very small
genomes, still elude our understanding of
their complicated life cycle, and because
miRNAs, thanks to their small genomic
size, are perfect candidates for storing
regulatory information in the small viral
genomes. (I should note that among viruses, herpesviruses are giants since their
genomes contain 100–200 protein coding
genes, which is, however, still two orders
of magnitude smaller than the number of
genes encoded in the human genome.)
Moreover, herpesviruses have a very interesting life cycle, consisting of lytic and
latent infections, which are very familiar in
the case of the best known representative
of human herpesviruses, the herpes simplex virus 1 (HSV-1): In the lytic infection,
the virus replicates, causing the cold sores,
and most of the viral genes are expressed;
as a result the virus becomes visible to the
immune system, which eventually eliminates the virus, except for a few virion particles that escape to the trigeminal ganglia,
where they establish latent infection. In the
latent infection, which can last for many
years, the virus is dormant and all but invisible to the immune system. When the
host becomes exposed to biological stress,
such as heat shock, sunburn, infection, etc.,
the virus, which depends on the well-being
of its host, decides that it is time to find
another host, and reactivates, starting the
lytic infection again. In the case of varicella zoster virus (VZV), the primary infection – varicella (chicken pox), was even
for a long time believed to be caused by a
different pathogen than the reactivated infection – zoster (shingles).
Interestingly, in all herpesviruses there
is one or at most a handful of so-called immediate-early genes, expression of which
alone can cause reactivation. In the lytic
expression cascade, these immediate-early
genes activate early genes, responsible for
DNA replication, which turns on the expression of the late, structural genes, comprising the majority of the genome.
Using our miRNA target prediction algorithms, we predicted that a viral miRNA
hcmv-miR-112-1 targets the immediate
early protein 1 (IE1) mRNA of the human
cytomegalovirus,[16] and that human hsamiR-200 miRNA family members target
the 3'UTR of the immediate early protein
2 (IE2) of this virus,[24] thus helping either
to maintain latency or to enter latency from
the lytic infection. These predictions were
confirmed experimentally by Eain Murphy,
Tom Shenk, and Christine O’Connor from
Princeton University and from the Lerner
Research Institute.[16,24] Although the repression by miRNAs is only partial, targeting the immediate-early genes appears
to be the optimal way to repress the lytic
expression cascade.
In another collaboration, with Didier
Trono’s Laboratory of Virology and
Genetics at EPFL, we identified a group of
genes playing a critical role in the maturation of red blood cells (erythropoiesis).
Building on preliminary miRNA and
mRNA expression data obtained in Didier
Trono’s laboratory, we predicted theoretically which candidate genes were the most
likely to be repressed by specific miRNAs.
These predictions included genes mediating mitophagy – a process involving
the elimination of mitochondria from red
blood cells and maximizing the cells’ ability to carry hemoglobin – and were confirmed experimentally in Trono’s laboratory.[25]
6. Conclusion
In conclusion, I have described several
simple, yet effective algorithms for predicting targets of both conserved[13,16,20,22] and
species-specific[13,15,20,22] miRNAs, with
binding sites both in the coding[20] and 3'
untranslated regions.[13,15,16,22] In all of our
algorithms we attempt to avoid empirical
parameters and training sets; the only indispensable parameters are the length and
location of the seed region (7 nt starting at
position 2 of the miRNA), which are widely accepted parameters in most precise
algorithms. We first select only the seed
matches that are partially accessible and/
or conserved in several species; the predictions with binding sites both in the coding
and 3’ untranslated regions are then ranked
according to the over-representation of the
conserved and/or accessible seed matches
relative to an appropriate random background. At present, we work on removing the remaining parameters, such as the
thresholds for conservation and accessibility, and on extending the algorithms to include the information from high throughput RNA expression experiments.
Acknowledgments
I would like to thank Ray Marín and
Miroslav Šulc, who worked in my group on the
theoretical and computational aspects of the
projects discussed here, Harlan Robins from
Fred Hutchinson Cancer Research Center and
Arnold Levine from the Institute for Advanced
Study for introducing me to the miRNAs and
herpesviruses, our experimental collaborators
Christine O’Connor and Eain Murphy from
the Lerner Research Institute at the Cleveland
Clinic, Tom Shenk from Princeton University,
and Isabelle Barde and Didier Trono from
EPFL. This work was supported by EPFL.
Finally I thank Ray MarÍn, Miroslav Šulc, and
Eduardo Zambrano for providing the figures.
Received: July 13, 2014
[1] B. P. Lewis, C. B. Burge, D. P. Bartel, Cell
2005, 120, 15.
[2] W. Filipowicz, S. N. Bhattacharyya, N.
Sonenberg, Nat. Rev. Genet. 2008, 9, 102.
[3] M. Selbach, B. Schwanhausser, N. Thierfelder,
Z. Fang, R. Khanin, N. Rajewsky, Nature 2008,
455, 58.
[4] D. Baek, J. Villen, C. Shin, F. Camargo, S.
Gygi, D. Bartel, Nature 2008, 455, 64.
[5] M. Hafner, M. Landthaler, L. Burger, M.
Khorshid, J. Hausser, P. Berninger, A.
Rothballer, M. Ascano Jr, A.-C. Jungkamp,
M. Munschauer, A. Ulrich, G. S. Wardle, S.
Dewell, M. Zavolan, T. Tuschl, Cell 2010, 141,
129.
[6] S. W. Chi, J. B. Zang, A. Mele, R. B. Darnell,
Nature 2009, 460, 479.
[7] M. Maragkakis, P. Alexiou, G. Papadopoulos,
M. Reczko, T. Dalamagas, G. Giannopoulos,
G. Goumas, E. Koukis, K. Kourtis, V. Simossis,
P. Sethupathy, T. Vergoulis, N. Koziris, T.
Sellis, P. Tsanakas, A. Hatzigeorgiou, BMC
Bioinformatics 2009, 10, 295.
[8] A. Krek, D. Grun, M. N. Poy, R. Wolf, L.
Rosenberg, E. J. Epstein, P. MacMenamin,
I. da Piedade, K. C. Gunsalus, M. Stoffel, N.
Rajewsky, Nat. Genet. 2005, 37, 495.
[9] R. C. Friedman, K. K.-H. Farh, C. B. Burge, D.
P. Bartel, Genome Res. 2009, 19, 92.
[10] A. Grimson, K. K.-H. Farh, W. K. Johnston,
P. Garrett-Engele, L. P. Lim, D. P. Bartel, Mol.
Cell 2007, 27, 91.
[11] D. Gaidatzis, E. van Nimwegen, J. Hausser, M.
Zavolan, BMC Bioinformatics 2007, 8, 69.
[12] H. Robins, W. H. Press, Proc. Natl. Acad. Sci.
USA 2005, 102, 15557.
[13] R. M. MarÍn, J. Vaníček, PloS one 2012, 7,
e32208.
[14] H. Robins, Y. Li, R. W. Padgett, Proc. Natl.
Acad. Sci. USA 2005, 102, 4006.
[15] R. M. MarÍn, J. Vaníček, Nucl. Acids Res. 2011,
39, 19.
[16] E. Murphy, J. Vaníček, H. Robins, T. Shenk, A.
J. Levine, Proc. Natl. Acad. Sci. USA 2008, 105,
5453.
[17] M. Kertesz, N. Iovino, U. Unnerstall, U. Gaul,
E. Segal, Nat. Genet. 2007, 39, 1278.
[18] A. Fuglsang, Biochem. Bioph. Res. Co. 2004,
316, 755.
[19] H. Robins, M. Krasnitz, H. Barak, A. J. Levine,
J. Bacteriol. 2005, 187, 8370.
[20] R. M. MarÍn, M. Sulc, J. Vaníček, RNA 2013,
19, 467.
[21] D. E. Knuth, ‘The Art of Computer Programming’, Vol. 2: ‘Seminumerical algorithms’, 3rd
ed., Addison-Wesley: Boston, 1997.
[22] R. M. MarÍn, F. Voellmy, T. von Erlach, J.
Vaníček, RNA 2012, 18, 1760.
[23] Y. Wan, K. Qu, Q. C. Zhang, R. A. Flynn, O.
Manor, Z. Ouyang, J. Zhang, R. C. Spitale, M.
P. Snyder, E. Segal, H. Y. Chang, Nature 2014,
505, 706.
[24] C. M. O’Connor, J. Vaníček, E. A. Murphy, J.
Virology 2014, 88, 5524.
[25] I. Barde, B. Rauwel, R. M. MarÍn-Florez, A.
Corsinotti, E. Laurenti, S. Verp, S. Offner, J.
Marquis, A. Kapopoulou, J. Vaníček, D. Trono,
Science 2013, 340, 350.